An estimation of the optimal Gaussian kernel parameter for support vector classification

Abstract

The selection of kernel function and its parameters has heavy influence on the generalization performance of support vector machine (SVM), and it becomes a focus on SVM researches. At present, there are not general rules to select an optimal kernel function for a given problem yet, alternatively, Gaussian and Polynomial kernels are commonly used for practice applications. Based on the relationship analysis of Gaussian kernel support vector machine and scale space theory, this paper proves the existence of a certain range of the parameter σ , within the range the generalization performance is good. An appropriate σ within the range can be achieved via dynamic evaluation as well. Simulation results demonstrate the feasibility and effectiveness of the presented approach. © 2008 Springer-Verlag Berlin Heidelberg.